ESMFold2 vs AlphaFold 3 vs Boltz-2
Maintained by the Talindrew team · Last updated
ESMFold2 (Biohub, May 2026, MIT licence) predicts all-atom complexes of proteins, nucleic acids and small molecules from a 6-billion-parameter protein language model, optionally with an MSA, and reports folding a 1,024-residue complex in about 16 seconds on an H100. AlphaFold 3 (Google DeepMind, 2024) predicts the same kinds of complexes from MSAs and templates and remains a reference for accuracy, but its weights are non-commercial. Boltz-2 (MIT) co-folds a protein with a ligand and also predicts binding affinity, which makes it the most direct fit for ranking compounds. The original ESMFold folds single protein chains in seconds and is the quickest first look.
Watch · 90 sec
Drug Discovery in 90 Seconds: Build and Run a Screen in Talindrew's AI Workspace
Chapters
How do they differ in what they take in?
ESMFold reads one protein sequence and nothing else. ESMFold2 reads proteins, DNA, RNA and small molecules, with an MSA optional (ESMFold2-Fast is single sequence only). AlphaFold 3 and Boltz-2 read the same complex types and use MSAs for best accuracy; Boltz-2 can also run without one.
Which one for which step of drug discovery?
Match the model to the question, not to the leaderboard.
- A quick structure of a monomer target to find pockets on: ESMFold.
- Protein-protein or antibody-antigen complexes: ESMFold2 or AlphaFold 3 (non-commercial use only).
- A protein with a candidate ligand, plus an affinity estimate to rank by: Boltz-2.
- A target with an experimental structure: skip prediction and bring the PDB entry.
Can I trust the confidence scores?
All four report pLDDT per residue (shown on a 0 to 1 scale in Talindrew), and the complex models add pTM and ipTM for interfaces. Low-confidence loops near a pocket make docking into it unreliable, whichever model produced them; check per-residue confidence around the pocket before trusting a docking score.
Structure prediction models for drug discovery (checked 2026-10-02)
| Model | Inputs | Ligands | Licence | In Talindrew |
|---|---|---|---|---|
| ESMFold | One protein chain, no MSA | No | MIT | Built in (fold stage default) |
| ESMFold2 | Proteins, DNA, RNA, small molecules; MSA optional | Yes | MIT | As your own model endpoint |
| AlphaFold 3 | Complexes, MSA and templates | Yes | Weights non-commercial | Your own licensed copy, or AlphaFold Server upload |
| Boltz-2 | Complexes, MSA optional | Yes, with affinity | MIT | Built in (fold and pose-check stages) |
Frequently asked questions
Is ESMFold2 better than AlphaFold 3?
Biohub reports ESMFold2 matching or beating AlphaFold 3 on protein-protein and antibody-antigen interface benchmarks, at much higher speed. Independent benchmarks were still appearing in late 2026, so test both on complexes like yours where licences allow.
Is ESMFold2 free for commercial use?
Yes. Biohub released ESMFold2's code and weights under the MIT licence, subject to its acceptable use policy.
References
- Lin Z, Akin H, Rao R, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379(6637): 1123–1130 (2023). https://doi.org/10.1126/science.ade2574
- Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630: 493–500 (2024). https://doi.org/10.1038/s41586-024-07487-w
- Passaro S, Corso G, Wohlwend J, et al. Boltz-2: Towards accurate and efficient binding affinity prediction. bioRxiv 2025.06.14.659707 (2025). https://doi.org/10.1101/2025.06.14.659707
- Biohub. ESMFold2 model card. https://huggingface.co/biohub/ESMFold2
Related
Design your ESMFold2 vs AlphaFold 3 vs Boltz-2 construct in Talindrew.
Free to use. No download required. Works in any modern browser.